SED-RCNN-BE: A SE-Dual channel RCNN network optimized binocular estimation model for automatic size estimation of free swimming fish in aquaculture

Abstract

Size of fish body is one of the most important morphological characteristics for cultured fish and plays a key role to enable effective management in aquaculture. Automatic, accurate and real-time monitoring of fish body size is therefore essential for improving aquaculture production. Computer vision facilitated the non-contact estimation of fish size and reduce labor consumption. While, current computer vision estimation techniques are still restricted in multi-scale size and multi-pose-variant of free-swimming fish, which limit the relevance application to aquaculture. Hence, we propose a novel method to facilitate automatic and accurate size estimation of free-swimming fish for growth monitoring of fish in aquaculture. This method comprised of an improved keypoint detection module based on Keypoint RCNN (SE-D-KP-RCNN) network and size estimation module based on CREStereo matching and spatial plane curve fitting. In the detection module, a squeeze and excitation Dual channel pooling layer (SE-D) is designed for the backbone network to improve the keypoint detection accuracy for free-swimming fish. In the estimation module, the global parallax map of binocular image is produced by CREStereo matching to obtain the 3D information of detection keypoint and build curve fitting of multi-pose variant fish body for height and length estimation. We implement and test our proposed method in different sizes and poses fish and the comprehensive experimental results show that the average accuracy of size estimation can reached 93.18% in the real aquaculture scene. This is important to help breeders to obtain underwater fish size information accurately and improve the efficiency for aquaculture management.